Signal processing method and device based on multi-physiological sound separation and artificial intelligence

Abstract

The invention provides a signal processing method and device based on multi-physiological sound separation and artificial intelligence, and relates to the field of physiological sound signal processing. The method comprises the following steps: acquiring multichannel physiological sound signals of the heart, the lung and the intestines and stomach by adopting a multichannel physiological sound acquisition technology, and pre-separating the physiological sound signals to improve the quality of original mixed physiological sound and the reliability of an acquisition process; inputting the pre-separated physiological sound components into the trained BP neural network model for component classification and source signal reconstruction to obtain separated heart sound signals, lung sound signals and borborygmus signals; and further respectively inputting the heart sound signal, the lung sound signal and the borborygmus signal into the corresponding classification network models for type classification. According to the invention, synchronous acquisition, separation and classification of physiological sounds of heart, lung and intestines and stomach parts by the same auscultation system are realized, and the accuracy of separation and classification of heart sounds, lung sounds and borborygmus is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of physiological sound signal processing, and in particular to a signal processing method and device based on multiple physiological sound separation and artificial intelligence. Background Art

[0002] The analysis of physiological sound signals plays an important role in the early screening and diagnosis of diseases. Physiological sound signals such as heart sounds, lung sounds, and bowel sounds contain rich pathological information. In recent years, with the development of artificial intelligence technology, many studies have begun to apply deep learning and other methods to the separation and diagnostic analysis of physiological sound mixed signals.

[0003] However, current research on physiological sound separation primarily focuses on single-channel heart and lung sound separation, which presents numerous shortcomings. For one thing, single-channel data captures limited information, making it difficult to fully capture its complex characteristics. Furthermore, it is significantly affected by the environment and user manipulation, making it difficult to obtain accurate physiological sound separation results when the collected heart and lung sound data is of poor quality. Furthermore, single-channel heart and lung sound separation often overlooks the interference of bowel sounds. In actual auscultation, heart sounds, lung sounds, and bowel sounds are intertwined. Failure to fully consider the influence of bowel sounds during signal separation can easily lead to misjudgment, impacting the accuracy of heart, lung, and bowel sound classification. Summary of the Invention

[0004] The purpose of the present invention is to provide a signal processing method and device based on multi-physiological sound separation and artificial intelligence to solve the problem mentioned in the above background technology that the existing technology is difficult to obtain accurate separation results of heart sounds, lung sounds and bowel sounds.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a signal processing method based on multi-physiological sound separation and artificial intelligence, the steps comprising: S1, obtaining multi-channel physiological sound signals of the heart, lungs and gastrointestinal tract, and performing pre-separation processing on the physiological sound signals to obtain physiological sound components; S2, establishing a BP neural network model, inputting the physiological sound components into the trained BP neural network model for component classification and reconstructing the source signal to obtain heart sound signals, lung sound signals and bowel sound signals; S3, establishing a classification network model, inputting the heart sound signal, the lung sound signal and the bowel sound signal into the corresponding classification network model respectively for type classification to obtain heart sound type, lung sound type and bowel sound type.

[0006] Optionally, the S1 step specifically includes: denoising the physiological sound signal; performing empirical mode decomposition on the denoised physiological sound signal to obtain intrinsic mode function components, and performing Hilbert transform on each of the intrinsic mode function components to calculate the instantaneous frequency; establishing an intrinsic mode function matrix component containing the instantaneous frequency, and analyzing it through an independent vector analysis algorithm to obtain a separation component matrix.

[0007] Optionally, the S2 step specifically includes: inputting the independent separated components in the separated component matrix into the trained BP neural network model for component classification, wherein the physiological sound components of the cardiac channel retain the heart sound components, the physiological sound components of the pulmonary channel retain the lung sound components, and the physiological sound components of the gastrointestinal channel retain the bowel sound components; converting the heart sound components, the lung sound components and the bowel sound components from spectral signals into time domain signals through inverse short-time Fourier transform to obtain heart sound signals, lung sound signals and bowel sound signals.

[0008] Optionally, the training steps of the BP neural network model include: S201, classifying and labeling the physiological sound components, and extracting the physiological sound features of the physiological sound components; S202, training the BP neural network model using the physiological sound features and the corresponding component labels as training samples; wherein the physiological sound features include: the power ratio of the 10Hz-50Hz frequency band to the 10Hz-2000Hz frequency band, the power ratio of the 50Hz-200Hz frequency band to the 10Hz-2 The power ratios of the 400Hz-1000Hz frequency band and the 10Hz-2000Hz frequency band are as follows: the power ratio of the 1000Hz-1200Hz frequency band and the 10Hz-2000Hz frequency band, the power ratio of the 1200Hz-2000Hz frequency band and the 10Hz-2000Hz frequency band, the standard deviation of the ventricular contraction cycle interval, and the peak frequency of the power spectral density.

[0009] Optionally, the training step of the BP neural network model also includes: S203, assigning the same initial weight to each training sample; S204, training the BP neural network weak classifier based on the weight of the current training sample, adjusting and calculating the weight of the weak classifier to minimize the classification error; S205, increasing the weight of the training sample with classification error, and reducing the weight of the training sample with classification correct; S206, judging whether the classification error of the current weak classifier is lower than a preset threshold; if not, returning to the S204 step to train the next weak classifier; if so, stopping the training and entering the S207 step; S207, combining all the weak classifiers and their weights into a strong classifier, and outputting the component classification result through the strong classifier.

[0010] On the other hand, the present invention also provides a signal processing device based on multi-physiological sound separation and artificial intelligence, including: a pre-separation module for acquiring multi-channel physiological sound signals of the heart, lungs and gastrointestinal tract, and performing pre-separation processing on the physiological sound signals to obtain physiological sound components; a component classification module for establishing a BP neural network model, inputting the physiological sound components into the trained BP neural network model for component classification and reconstructing the source signal to obtain heart sound signals, lung sound signals and bowel sound signals; a type classification module for establishing a classification network model, respectively inputting the heart sound signal, the lung sound signal and the bowel sound signal into the corresponding classification network model for type classification to obtain heart sound type, lung sound type and bowel sound type.

[0011] Optionally, the type classification module specifically includes: a heart sound classification module, used to input the heart sound signal into the heart sound classification network model for type classification, and determine the heart sound type by weighted fusion, wherein the heart sound types include normal heart sounds, heart murmurs, extra heart sounds and premature beats; a lung sound classification module, used to input the lung sound signal into the lung sound classification network model for type classification, and determine the lung sound type by weighted fusion, wherein the lung sound types include normal lung sounds, whistle sounds, moist rales, weakened alveolar breath sounds, and enhanced alveolar breath sounds; a bowel sound classification module, used to input the bowel sound signal into the bowel sound classification network model for type classification, and determine the bowel sound type by weighted fusion, wherein the bowel sound types include normal bowel sounds, active bowel sounds, hyperactive bowel sounds, weakened bowel sounds, and disappeared bowel sounds.

[0012] Optionally, the heart sound classification network model includes: a heart sound feature extraction module, a convolution block, a maximum pooling layer, a random dropout layer, a channel attention layer, a self-attention layer, an LSTM module, and a first fully connected layer connected in sequence; the heart sound feature extraction module extracts several heart sound features, and splices and fuses them to form a multidimensional feature matrix as input data of the heart sound classification network model, wherein the heart sound features include Mel-frequency cepstral coefficients, Mel-spectrograms, spectral contrast, and time-frequency features of heart sound signals extracted by short-time Fourier transform.

[0013] Optionally, the lung sound classification network model includes: a lung sound feature extraction module, a multi-scale convolution block module, an intra-scale attention module, an inter-scale attention module, a second fully connected layer, an LSTM module, and a third fully connected layer connected in sequence; the lung sound feature extraction module extracts several lung sound features and splices them to form a multi-dimensional feature matrix as input data of the lung sound classification network model, wherein the lung sound features include Mel-frequency cepstral coefficients, Mel-frequency spectrum, zero-crossing rate, and spectral centroid.

[0014] Optionally, the bowel sound classification network model includes: a bowel sound feature extraction module, a ResNet50 network module, and a multi-head attention mechanism module connected in sequence; the bowel sound feature extraction module extracts several bowel sound features and splices them into a multidimensional feature matrix as input data for the lung sound classification network model, wherein the bowel sound features include Mel-frequency cepstral coefficients and Mel-spectrograms.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention adopts multi-channel physiological sound acquisition technology to obtain multi-channel physiological sound signals of the heart, lungs and gastrointestinal tract, and then pre-separates the physiological sound signals to improve the quality of the collected original mixed physiological sounds; by inputting the physiological sound components after pre-separation into the trained BP neural network model for component classification and reconstructing the source signal, the separated heart sound signals, lung sound signals and bowel sound signals are obtained; by inputting the heart sound signals, lung sound signals and bowel sound signals into the corresponding classification network model for type classification, the heart sound type, lung sound type and bowel sound type are obtained. This method combines multi-channel physiological sound acquisition technology with artificial intelligence technology to realize the synchronous acquisition, separation and classification of physiological sounds of the heart, lungs and gastrointestinal tract by the same auscultation system, significantly improving the accuracy of separation and classification of heart sounds, lung sounds and bowel sounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of the process steps of the present invention.

[0017] Figure 2 It is a schematic diagram of the overall system flow of the present invention.

[0018] Figure 3 The figure is a schematic diagram of the process of separating heart sounds, lung sounds and bowel sounds of the present invention.

[0019] Figure 4 This is a schematic diagram of the BP neural network model training of the present invention.

[0020] Figure 5 This is a schematic diagram of the heart sound classification network model training of the present invention.

[0021] Figure 6 This is a schematic diagram of the lung sound classification network model training of the present invention.

[0022] Figure 7 This is a schematic diagram of the bowel sound classification network model training of the present invention.

[0023] Figure 8 It is a schematic diagram of the structure of the device of the present invention.

[0024] In the figure: 10-pre-separation module, 20-component classification module, 30-type classification module. DETAILED DESCRIPTION

[0025] The following will provide a clear and complete description of the solutions of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0026] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0027] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of this application refers to the presence of features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or wireless couplings. The term "and / or" used herein includes all or any units and all combinations of one or more associated listed items.

[0028] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0029] It should be understood that the sequence numbers and sizes of the steps in this embodiment do not imply the order of execution. The order of execution of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of this application.

[0030] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] Please refer to Figure 1-Figure 2 The present invention provides a signal processing method based on multi-physiological sound separation and artificial intelligence, comprising the following steps: S1, acquiring multi-channel physiological sound signals of the heart, lungs, and stomach, and performing pre-separation processing on the physiological sound signals to obtain physiological sound components.

[0032] Specifically, the user places thirteen HKY-6E sensors on the corresponding body parts according to the system instructions, remains still, and then chooses to start diagnosis; the propagation of physiological sound signals will cause instantaneous pressure changes and tiny deformations on the surface of the human skin. The HKY-6E sensor adopts a capacitive working principle, the core of which is based on the characteristics of capacitance; the sensor consists of a fixed electrode and a movable electrode. When subjected to external pressure or deformation, the distance between the movable electrode and the fixed electrode, the relative area or the dielectric constant of the medium between the electrodes will change. The changes in these parameters will cause the capacitance value to change accordingly, thereby converting the mechanical vibration signal into a measurable electrical signal; the frequency response range of HKY-6E is usually 10-3000Hz, which completely covers the frequency characteristics of heart sounds, lung sounds and bowel sounds, so it can effectively record and analyze these three signals.

[0033] Optionally, six sensors are used in the lung passage area: The right upper anterior (RUA) is located in the right upper lobe region, corresponding to the anterior surfaces of the 2nd to 4th ribs; the left upper anterior (LUA) is located in the left upper lobe region, corresponding to the anterior surfaces of the 4th to 6th ribs; the right middle anterior (RMA) is located in the right middle lobe region, corresponding to the anterior surfaces of the 4th to 6th ribs; the left middle anterior (LMA) is located in the left middle lobe region, corresponding to the anterior surfaces of the 4th to 6th ribs; the right lower anterior (RLA) is located in the right lower lobe region, corresponding to the anterior surfaces of the 6th to 8th ribs, at a 45° angle downward from the nipple; the left lower anterior (LLA) is located in the left lower lobe region, corresponding to the anterior surfaces of the 6th to 8th ribs, at a 45° angle downward from the nipple.

[0034] Optionally, the heart channel area uses four sensors: The apex (Apex, A) is located at the 5th intercostal space, midclavicular line, which is the best position for auscultation of the mitral valve, usually used to auscultate the first heart sound (S1) and the murmur of mitral stenosis or regurgitation; the right upper sternal border (RUSB) is located at the 2nd intercostal space on the right side of the sternum, which is the main area for auscultation of the aortic valve, usually used to auscultate the murmur of aortic stenosis or regurgitation; the left upper sternal border (LUSB) is located at the 2nd intercostal space on the left side of the sternum, which is the main area for auscultation of the pulmonary valve, usually used to auscultate the murmur of pulmonary stenosis or regurgitation; the left lower sternal border (LLSB) is located at the 4th and 5th intercostal spaces on the left side of the sternum, which is the main area for auscultation of the tricuspid valve. The main area of ​​auscultation for tricuspid valve auscultation is usually used to auscultate the murmur of tricuspid stenosis or regurgitation.

[0035] Optionally, the gastrointestinal passage area uses three sensors: The posterior abdominal area (PAA) is the auscultation area on the posterior side of the abdomen, used to listen to bowel sounds, especially near the navel; the left anterior abdominal area (LAAA) is the auscultation area on the left anterior side of the abdomen, used to listen to bowel sounds in the spleen area; the right anterior abdominal area (RAAA) is the auscultation area on the right anterior side of the abdomen, used to listen to bowel sounds in the liver area.

[0036] S2, establishing a BP neural network model, inputting the physiological sound components into the trained BP neural network model to perform component classification and reconstruct the source signal to obtain heart sound signals, lung sound signals and bowel sound signals.

[0037] Specifically, the physiological sound separation process of the present application combines the Hilbert-Huang Transform (HHT) with the Independent Vector Analysis (IVA) blind source separation algorithm to perform pre-separation processing on the physiological sounds, and further separates the heart sound components, lung sound components, bowel sound components and noise components through the BP neural network model. Compared with the independent vector analysis blind source separation algorithm, the accuracy of the separated signal is greatly improved.

[0038] S3, establishing a classification network model, inputting the heart sound signal, the lung sound signal, and the bowel sound signal into the corresponding classification network model respectively for type classification, to obtain the heart sound type, lung sound type, and bowel sound type.

[0039] Specifically, the four-channel heart sound data obtained in the separation stage are input into the trained heart sound classification model to determine the final heart sound classification result, the six-channel lung sound data obtained in the separation stage are input into the trained lung sound classification model to determine the final lung sound classification result, and the three-channel bowel sound data obtained in the separation stage are input into the trained bowel sound classification model to determine the final bowel sound classification result, which greatly improves the accuracy and reliability of heart sound, lung sound and bowel sound classification.

[0040] It can be understood that the present application adopts multi-channel physiological sound separation technology to obtain multi-channel physiological sound signals of the heart, lungs and gastrointestinal tract, and pre-separates the physiological sound signals, so that the quality of the original mixed physiological sounds collected is improved to a certain extent, and the collection process is more reliable; by inputting the physiological sound components after pre-separation processing into the trained BP neural network model for component classification and reconstructing the source signal, heart sound signals, lung sound signals and bowel sound signals are obtained, which can improve the accuracy and reliability of physiological sound signal separation; by inputting the heart sound signal, the lung sound signal and the bowel sound signal into the corresponding classification network model for type classification, the heart sound type, lung sound type and bowel sound type are obtained, which can improve the accuracy and reliability of heart sound, lung sound and bowel sound classification.

[0041] Please refer to Figure 3 In some embodiments, the S1 step specifically includes: denoising the physiological sound signal; performing empirical mode decomposition on the denoised physiological sound signal to obtain intrinsic mode function components, and performing Hilbert transform on each of the intrinsic mode function components to calculate the instantaneous frequency; establishing an intrinsic mode function matrix component containing the instantaneous frequency, and analyzing it through an independent vector analysis algorithm to obtain a separation component matrix.

[0042] Specifically, the data of the thirteen channels are subjected to Wiener filtering-spectral subtraction noise reduction, first processed by short-time Fourier transform STFT, in which the Hanning window function Window length is 2048, overlap length is 1024, and through this transformation, the mixed physiological sound signal in the time domain is recorded as The spectrum of the noisy signal converted to the frequency domain is denoted as , further using spectral subtraction, the calculation formula is: , where is a preliminary estimate of the spectrum of the pure physiological sound signal, is the over-reduction factor and has a value of 2. is the noise spectrum.

[0043] Furthermore, Wiener filtering is used, and its calculation formula is: , where To estimate the spectrum, the is the power spectrum of the pure signal, is the power spectrum of the noise; based on the result of spectral subtraction, it is further optimized with the help of Wiener filtering, taking full account of the statistical characteristics of the signal and noise to obtain a more accurate estimated spectrum, and then performing inverse short-time Fourier transform to obtain the time domain signal after noise reduction The signal will be used as the input of the next Hilbert-Huang transform. The Wiener filter-spectral subtraction noise reduction can effectively eliminate the interference in the physiological sound signal and greatly improve the signal quality.

[0044] Furthermore, the Hilbert-Huang transform (HHT) is used, which includes two parts: empirical mode decomposition (EMD) and Hilbert transform. The empirical mode decomposition calculation formula is: , where For the The signal obtained by the second screening is For the The signal obtained by the second screening is is the mean of the upper and lower envelopes; for the signal after noise reduction Perform empirical mode decomposition and decompose it into several intrinsic mode function IMF components and a residual function The standard deviation criterion is used as the stopping condition. When the standard deviation of the IMF component obtained by two consecutive screenings is less than the threshold value of 0.001, the decomposition of the IMF component is considered to be completed. These IMF components cover the characteristics of different time scales of the signal.

[0045] Furthermore, the instantaneous frequency of the IMF component is calculated using the Hilbert transform. Each IMF component is subjected to a Hilbert transform to obtain an analytical signal, and then the instantaneous frequency is calculated. These IMF matrix components containing the instantaneous frequency information will serve as the input of the independent vector analysis (IVA).

[0046] Furthermore, after the Hilbert-Huang transform (HHT), the obtained IMF components carry rich instantaneous frequency information of the physiological sound signal. These IMF components are arranged in sequence to construct an IMF matrix, which serves as the multi-channel input signal of the independent vector analysis (IVA) algorithm. Each IMF component occupies a specific column in the matrix, and different columns represent different "channels". These channel signals are a mixture of relevant characteristic components of heart sounds, lung sounds, bowel sounds and noise. The IVA algorithm updates the separation matrix through continuous iteration. , when the convergence condition is met, that is, the separation matrix between two iterations is satisfied When the change of is less than the convergence threshold, the threshold Set to , and obtain the final separation matrix , the input multi-channel signal With the separation matrix Multiplying, the calculation formula is: , where is the component matrix after separation, is the separation matrix, For a multi-channel signal, each row in this matrix represents an independent separation component. Compared with the method of using only Hilbert-Huang transform HHT or independent vector analysis IVA algorithm for separation, the present application effectively solves the separation problem of complex physiological sound signals by combining Hilbert-Huang transform HHT with independent vector analysis IVA algorithm.

[0047] Please refer to Figure 3 In some embodiments, the S2 step specifically includes: inputting the independent separated components in the separated component matrix into the trained BP neural network model for component classification, wherein the physiological sound components of the cardiac channel retain the heart sound components, the physiological sound components of the pulmonary channel retain the lung sound components, and the physiological sound components of the gastrointestinal channel retain the bowel sound components; converting the heart sound components, the lung sound components, and the bowel sound components from spectral signals into time domain signals through inverse short-time Fourier transform to obtain heart sound signals, lung sound signals, and bowel sound signals.

[0048] Specifically, by inputting the independent separated components into the trained BP neural network model, the classification of the separated independent components can be achieved, including heart sound components, lung sound components, bowel sound components, and noise components; only heart sound components are retained in the four-channel data of the heart, only lung sound components are retained in the six-channel data of the lungs, and only bowel sound components are retained in the three-channel data of the gastrointestinal tract; using the inverse short-time Fourier transform (ISTFT), the heart sound components, lung sound components, and bowel sound components classified by the BP neural network are converted from spectrum signals back to time domain signals to obtain four channels of pure heart sound data, six channels of pure lung sound data, and three channels of pure bowel sound data after separation; the present application further processes the pre-separated signals using the BP neural network model, which can more fully separate the heart sound components, lung sound components, and bowel sound components, effectively avoiding the influence of auscultation misjudgment caused by the interweaving of heart sound, lung sound, and bowel sound auscultation signals during the auscultation process, making it difficult to fully separate the heart sound components, lung sound components, and bowel sound components, thereby greatly improving the accuracy and reliability of signal separation.

[0049] Please refer to Figure 4 In some embodiments, the training step of the BP neural network model includes: S201, classifying and labeling the physiological sound components after the pre-separation process, and extracting the physiological sound features of the physiological sound components after the pre-separation process; S202, using the physiological sound features and the corresponding component labels as training samples to train the BP neural network model; wherein the physiological sound features include: the power ratio of the 10Hz-50Hz frequency band to the 10Hz-2000Hz frequency band, the 50Hz-200Hz The power ratio of the z band to the 10Hz-2000Hz band, the power ratio of the 200Hz-400Hz band to the 10Hz-2000Hz band, the power ratio of the 400Hz-1000Hz band to the 10Hz-2000Hz band, the power ratio of the 1000Hz-1200Hz band to the 10Hz-2000Hz band, the power ratio of the 1200Hz-2000Hz band to the 10Hz-2000Hz band, the standard deviation of the ventricular contraction cycle interval, and the peak frequency of the power spectral density.

[0050] Specifically, eight physiological sound feature matrices of heart sound component, lung sound component, bowel sound component, and noise component are extracted from the physiological sound independent component training data, and the corresponding component labels are used as training samples for the BP neural network. The eight physiological sound features include: The power ratio of the 10Hz-50Hz frequency band to the 10Hz-2000Hz frequency band. 10Hz-50Hz is considered the dominant frequency band for bowel sounds. The ratio of the power of this frequency band to the power of the total 10Hz-2000Hz frequency band can reflect the proportion of bowel sounds in the entire signal energy. If this ratio is high, it means that the bowel sound component in the signal is relatively strong, which may indicate that intestinal peristalsis is relatively active, which may be related to factors such as the body's digestive state and eating habits. For example, shortly after eating, intestinal peristalsis will accelerate and bowel sounds will also increase accordingly. At this time, the ratio may increase. Conversely, if the ratio is low, it indicates that the bowel sound component is weak, which may be due to the intestine being in a relatively quiet state or being interfered with by other physiological sounds, such as heart sounds, lung sounds, or noise.

[0051] The power ratio of the 50Hz-200Hz frequency band to the 10Hz-2000Hz frequency band (50Hz-200Hz is considered the dominant frequency band for heart sounds). The ratio of the power of this frequency band to the total power of the 10Hz-2000Hz frequency band reflects the contribution of heart sounds to the overall signal energy. A high ratio indicates a relatively strong heart sound component in the signal; a low ratio indicates a weak heart sound component or interference from other physiological sounds, such as lung sounds, bowel sounds, or noise. The intensity and characteristics of heart sounds are closely related to the functional state of the heart, and monitoring this ratio can help determine the health of the heart.

[0052] The power ratio of the 200Hz-400Hz frequency band to the 10Hz-2000Hz frequency band. The 200Hz-400Hz frequency band is a mixed frequency band for heart and lung sounds. Calculating the power ratio of this frequency band to the total frequency band can reveal the energy contribution of the mixed components of heart and lung sounds to the overall signal. Changes in this ratio reflect the relative intensity relationship between heart and lung sounds and are valuable for assessing cardiopulmonary function. For example, when cardiopulmonary function is abnormal, the intensity of heart and lung sounds may change, causing fluctuations in this ratio.

[0053] The power ratio of the 400Hz-1000Hz frequency band to the 10Hz-2000Hz frequency band. The 400Hz-1000Hz frequency band is the dominant frequency band for lung sounds. The ratio of the power of this frequency band to the total frequency band power reflects the intensity of lung sounds in the signal. A higher ratio indicates that lung sounds are the main component of the signal, which may be related to normal lung respiratory activity or certain pathological conditions. For example, during an attack of a lung disease such as pneumonia or asthma, the characteristics and intensity of lung sounds will change significantly, and this ratio will also change accordingly.

[0054] The power ratio of the 1000Hz-1200Hz frequency band to the 10Hz-2000Hz frequency band. The 1000Hz-1200Hz frequency band is a mixed frequency band of lung sounds and noise. Calculating the power ratio in this frequency band can assess the degree of noise interference with the lung sound signal. An abnormally high ratio may indicate that the signal is subject to excessive noise contamination and requires further denoising. Bowel sounds may also make a slight, but relatively small, contribution in this frequency band. By monitoring this ratio, noise issues in the signal can be promptly identified and signal quality improved.

[0055] The power ratio of the 1200Hz-2000Hz frequency band to the 10Hz-2000Hz frequency band is considered the dominant frequency band for noise. The ratio of the power in this frequency band to the total frequency band power can intuitively reflect the proportion of noise energy in the signal. By monitoring this ratio, the signal quality can be judged. When the ratio is too high, measures may be needed to improve the signal-to-noise ratio. Bowel sounds typically have very low energy in this frequency band, which has little impact on this ratio. When calculating the ratio of power to total power within different frequency ranges, the signal's power spectral density is first estimated. Then, the power within each frequency band and the total frequency band power are calculated based on the frequency band range. Finally, the ratio is calculated. These ratios are used as features and input into the classification model, which can help the model more comprehensively understand the characteristics of different types of physiological sound signals, thereby better distinguishing different physiological sound signals such as heart sounds, lung sounds, and bowel sounds.

[0056] The standard deviation of the S1-S1 interval. S1 is the first heart sound, representing the onset of ventricular contraction. The standard deviation of the S1-S1 interval is a characteristic related to periodicity. Under normal circumstances, the heartbeat exhibits a certain degree of periodicity, and the S1-S1 interval is relatively stable. However, when the heart develops a disease or abnormality, this periodicity may be disrupted, and the S1-S1 interval may fluctuate. By calculating the standard deviation of the S1-S1 interval, the degree of this fluctuation can be quantified.

[0057] The peak frequency of power spectral density (PSD) is an effective tool for describing the energy distribution of a signal at different frequencies. Different types of physiological sounds exhibit different characteristics in power spectral density due to their different physiological mechanisms and signal characteristics.

[0058] It is difficult to fully describe the characteristics of physiological sound signals using only a single feature, resulting in inaccurate classification; the multi-feature comprehensive extraction method describes physiological sound signals from different angles, namely spectrum, time domain, periodicity, etc., which solves the problem of insufficient information from a single feature and enables the BP neural network model to have a more comprehensive understanding of the signal characteristics.

[0059] Please refer to Figure 4In some embodiments, the training step of the BP neural network model further includes: S203, assigning the same initial weight to each training sample; S204, training the BP neural network weak classifier based on the weight of the current training sample, adjusting and calculating the weight of the weak classifier to minimize the classification error; S205, increasing the weight of the training sample with classification error, and reducing the weight of the training sample with classification correctness; S206, judging whether the classification error of the current weak classifier is lower than a preset threshold; if not, returning to the step S24 to train the next weak classifier; if so, stopping the training and entering the step S207; S207, combining all the weak classifiers and their weights into a strong classifier, and outputting the component classification result through the strong classifier.

[0060] Specifically, the same initial weight is assigned to each training sample to ensure that all samples have the same importance in the initial stage.

[0061] Furthermore, a strong classifier is constructed by iteratively training multiple BP neural network classifiers. The specific steps are as follows: Update BP neural network weights: Train a BP neural network classifier based on the current sample weights. The BP neural network adjusts the weights through the backpropagation algorithm to minimize the classification error. Calculate the weight of the weak classifier: Calculate the error based on the classification result of the BP neural network, and calculate its weight based on the error. The smaller the classification error, the greater the weight of the BP neural network; Adjust sample weights: Increase the weights of misclassified samples and reduce the weights of correctly classified samples, so that the model pays more attention to difficult-to-classify samples in subsequent iterations; Checking the prediction error: In each iteration, the prediction error of the current weak classifier is checked to see if it is below a preset threshold. If the error is above the threshold, training continues with the next weak classifier. If the error is below the threshold, training stops and the strong classifier is integrated.

[0062] Furthermore, all BP neural network classifiers and their weights are combined into a strong classifier. Through a weighted voting mechanism, the strong classifier integrates the prediction results of multiple BP neural networks and outputs the final classification results of the separated independent components, thereby improving the classification accuracy of the heart sound component, lung sound component, and bowel sound component. Finally, the classifier output is converted into a probability using the Softmax function, representing the probability distribution of the input separated components belonging to heart sound, lung sound, bowel sound, and noise. The model is evaluated and optimized using appropriate evaluation metrics, and the optimal model is saved as the BP neural network classification model for the heart sound component, lung sound component, bowel sound component, and noise component. Single BP neural network classifiers are prone to overfitting and lack accuracy and robustness in classifying complex physiological sound signals. The integrated BP neural network constructs a strong classifier by iteratively training multiple classifiers, which increases the robustness of the model, makes the model more focused on difficult-to-classify samples, and improves classification accuracy.

[0063] Please refer to Figure 8 On the other hand, the present invention also provides a signal processing device based on multi-physiological sound separation and artificial intelligence, including: a pre-separation module 10, used to obtain multi-channel physiological sound signals of the heart, lungs and gastrointestinal tract, and perform pre-separation processing on the physiological sound signals; a component classification module 20, used to establish a BP neural network model, input the physiological sound components after pre-separation processing into the trained BP neural network model for component classification and reconstruct the source signal to obtain heart sound signals, lung sound signals and bowel sound signals; a type classification module 30, used to establish a classification network model, input the heart sound signal, the lung sound signal and the bowel sound signal into the corresponding classification network model for type classification, and obtain heart sound type, lung sound type and bowel sound type.

[0064] Specifically, the present application adopts multi-channel physiological sound separation technology to obtain multi-channel physiological sound signals of the heart, lungs and gastrointestinal tract, and pre-separates the physiological sound signals, so that the quality of the collected original mixed physiological sounds is improved to a certain extent, and the collection process is more reliable; by inputting the physiological sound components into the trained BP neural network model for component classification and reconstructing the source signal, heart sound signals, lung sound signals and bowel sound signals are obtained, which can improve the accuracy and reliability of physiological sound signal separation; by inputting the heart sound signal, the lung sound signal and the bowel sound signal into the corresponding classification network model for type classification, the heart sound type, lung sound type and bowel sound type are obtained, which can improve the accuracy and reliability of heart sound, lung sound and bowel sound classification.

[0065] In some embodiments, the type classification module specifically includes: a heart sound classification module, which is used to input the heart sound signal into the heart sound classification network model for type classification, and use a weighted fusion method to determine the heart sound type, wherein the heart sound types include normal heart sounds, heart murmurs, extra heart sounds and premature beats; a lung sound classification module, which is used to input the lung sound signal into the lung sound classification network model for type classification, and use a weighted fusion method to determine the lung sound type, wherein the lung sound types include normal lung sounds, whistle sounds, moist rales, weakened alveolar breath sounds, and enhanced alveolar breath sounds; a bowel sound classification module, which is used to input the bowel sound signal into the bowel sound classification network model for type classification, and use a weighted fusion method to determine the bowel sound type, wherein the bowel sound types include normal bowel sounds, active bowel sounds, hyperactive bowel sounds, weakened bowel sounds, and disappeared bowel sounds.

[0066] Specifically, in heart sound diagnosis, the four-channel heart sound data obtained in the separation stage is input into the trained heart sound classification model. A weighted fusion method is used to determine the final classification result. The weights of the data of each channel are set to the same value. The fused probability vector P{fusion}=[p{fusion1},p{fusion2},p{fusion3},p{fusion4}] is obtained. Subsequently, the category with the largest probability value in the vector is selected as the final classification result of the heart sound data; that is, one of the four categories of normal heart sound, heart murmur, extra heart sound, and premature beat, where: Normal heart sounds: Normal heart sounds mainly include the first heart sound S1 and the second heart sound S2; S1 marks the beginning of ventricular contraction, which is caused by the closure of the atrioventricular valve and the impact of blood flow on the ventricular wall; S2 marks the beginning of ventricular diastole, which is produced by the closure of the semilunar valve; normal heart sounds are clear and regular in rhythm, S1 has a lower pitch and a longer duration, S2 has a higher pitch and a shorter duration, and the interval between the two is regular, indicating that the heart structure and function are normal, the opening and closing activities of each valve are coordinated, and the hemodynamics are stable.

[0067] Heart murmur: It is caused by accelerated blood flow, stenosis of the valve orifice, valve insufficiency, abnormal blood flow channels, floating objects in the heart cavity or enlargement of the blood vessel cavity, which causes the blood flow to change from laminar flow to turbulent flow, and then forms a vortex, which hits the heart wall, valves, chordae tendineae or the wall of large blood vessels to produce vibration and form a sound; the intensity, nature (such as blowing, rumbling, machine-like, etc.), period (such as systole, diastole or continuity, etc.), conduction direction, etc. of the murmur vary; for example, the murmur of mitral regurgitation is mostly a systolic blowing murmur, which can be transmitted to the left armpit.

[0068] Murmurs of different natures and locations indicate different heart diseases, such as congenital heart disease, valvular disease, etc., which are of great significance for the diagnosis and assessment of the disease.

[0069] Extra heart sounds: They are additional heart sounds heard beyond the normal S1 and S2. According to the period of their appearance, they can be divided into diastolic extra heart sounds (such as gallop rhythm, opening valve sound, etc.) and systolic extra heart sounds (such as early systolic ejection sound, mid-to-late systolic click sound, etc.). Their occurrence is related to the vibration or abnormal activity of structures such as ventricular walls, valves, and chordae tendineae. The time of appearance, tone, and intensity of extra heart sounds have their own characteristics. For example, the early diastolic gallop rhythm is a low-key, additional third heart sound, which forms a sound similar to a horse running with S1 and S2, indicating that the myocardium is severely damaged. The appearance of extra heart sounds often indicates that there are serious heart diseases, such as heart failure, myocardial infarction, etc., which is of great value in judging the severity and prognosis of the disease.

[0070] Premature beats: also known as premature contractions, refer to the premature excitation of ectopic pacemakers originating outside the sinoatrial node, causing premature heart beats; they can be divided into atrial premature beats, junctional premature beats and ventricular premature beats. Their occurrence is related to abnormal automaticity of myocardial cells, reentrant excitation and other factors.

[0071] A premature heartbeat occurs one heartbeat prematurely, following a long compensatory pause, compared to a normal heart rhythm. The first heart sound is often enhanced, while the second heart sound is often weakened or absent. Occasional premature beats can occur in healthy individuals and are often associated with factors such as stress, overwork, and alcohol consumption. Frequent premature beats are common in patients with organic heart disease, such as coronary artery disease and cardiomyopathy, and can affect heart function and even lead to serious arrhythmias.

[0072] In lung sound diagnosis, the six-channel lung sound data obtained in the separation stage is input into the trained lung sound classification model. A weighted fusion approach is used to determine the final classification result. The weights of the data in each channel are set to the same value, resulting in a fused probability vector P{fusion}=[p{fusion1}, p{fusion2}, p{fusion3}, p{fusion4}, p{fusion5}]. Subsequently, the category with the largest probability value in the vector is selected as the final classification result of the lung sound data, namely, one of the five categories: normal lung sounds, whistles, crackles, weakened or absent alveolar breath sounds, and enhanced alveolar breath sounds. Normal lung sounds: appear as sighing or soft blowing "fu-fu" sounds. The sound is stronger, higher-pitched and longer-lasting during inhalation; the sound is weaker, lower-pitched and shorter-lasting during exhalation. Normal lung sounds indicate that the structure and function of the lungs are basically normal, gas flows smoothly in and out of the respiratory tract and alveoli, and the elasticity of the lung tissue and the patency of the airways are in good condition.

[0073] Whistling: A high-pitched, dry rales, similar to the sound of a whistle, are high-pitched, long-lasting, and musical. Their intensity and nature vary easily, and their location is not constant, with the number of rales increasing or decreasing significantly within a moment. They may indicate conditions such as bronchial asthma, chronic obstructive pulmonary disease (COPD), or foreign bodies in the trachea or bronchi.

[0074] Wet rales: They are the sounds produced by the bursting of bubbles formed by the passage of gas through secretions in the respiratory tract, such as sputum, exudate, blood, etc., during inhalation, so they are also called blister sounds. They are characterized by being intermittent and short-lived, often appearing multiple times in a row at a time. They are more obvious during inhalation or at the end of inhalation, and sometimes appear in the early stages of exhalation. Their location is relatively constant and their nature is not easy to change. Medium and small wet rales may exist at the same time, and may be alleviated or disappear after coughing. They may correspond to diseases such as pneumonia, bronchiectasis, and heart failure.

[0075] Weakened alveolar breath sounds: Compared with normal alveolar breath sounds, they are less intense, have lower pitch, and are shorter in duration. These may correspond to conditions such as restricted thoracic movement (e.g., chest pain, rib fractures), respiratory muscle diseases (e.g., myasthenia gravis), bronchial obstruction (e.g., chronic bronchitis, bronchial stenosis), and compressive atelectasis (e.g., pleural effusion, pneumothorax).

[0076] Enhanced alveolar breath sounds: Compared with normal alveolar breath sounds, the sound is stronger, the pitch is higher, and the duration is longer. The corresponding diseases may be: common in exercise, fever, hypermetabolism, anemia, acidosis, etc. At this time, the body's oxygen demand increases, causing deep and rapid breathing.

[0077] In bowel sound diagnosis, the three-channel bowel sound data obtained in the separation stage is input into the trained bowel sound classification model. A weighted fusion method is used to determine the final classification result. The weights of the data of each channel are set to the same value. The fused probability vector P{fusion} = [p{fusion1}, p{fusion2}, p{fusion3}, p{fusion4}, p{fusion5}] is obtained. Subsequently, the category with the largest probability value in the vector is selected as the final classification result of the bowel sound data, that is, one of the five categories: normal bowel sounds, active bowel sounds, hyperactive bowel sounds, hyperactive bowel sounds, and absent bowel sounds. Normal bowel sounds: Under normal circumstances, bowel sounds occur about 4-5 times per minute. The sound and tone are relatively even and gentle, similar to a "gurgling" sound in the stomach. This is the sound produced by the flow of gas and liquid in the intestine when the intestine moves, indicating that the intestine is in a normal digestive activity state.

[0078] Active bowel sounds: The frequency of bowel sounds increases, reaching over 10 times per minute, but the pitch is not particularly high. This is commonly seen in conditions such as acute gastroenteritis, after taking laxatives, or with severe gastrointestinal bleeding. For example, in acute gastroenteritis, the gastrointestinal mucosa is irritated, and peristalsis accelerates, resulting in active bowel sounds. Patients may also experience symptoms such as abdominal pain and diarrhea.

[0079] Hyperactive bowel sounds: Not only does the frequency of bowel sounds increase, but the tone is also loud and high-pitched, and may even sound like tinkling or metallic sounds. This is more common in mechanical intestinal obstruction. During intestinal obstruction, the intestine is blocked, and the passage of gas and liquid is obstructed. The intestinal tube moves strongly in an attempt to overcome the obstruction, resulting in hyperactive bowel sounds. At this time, patients often have typical symptoms of intestinal obstruction such as abdominal pain, vomiting, abdominal distension, and cessation of flatulence and defecation.

[0080] Weakened bowel sounds: Bowel sounds are significantly less than normal, and are heard only once every few minutes. This is common in senile constipation, peritonitis, electrolyte imbalance (such as hypokalemia), and low gastrointestinal motility. For example, in peritonitis, inflammatory stimulation leads to inhibition of intestinal wall muscle movement, which weakens bowel sounds.

[0081] Loss of bowel sounds: No bowel sounds are heard after 3-5 minutes of continuous auscultation, and there are still no bowel sounds when tapping or scratching the abdomen with fingers. This is called disappearance of bowel sounds. It is usually seen in acute peritonitis or paralytic intestinal obstruction. In these severe cases, intestinal peristalsis completely stops, and therefore no bowel sounds can be heard.

[0082] Please refer to Figure 5 In some embodiments, the heart sound classification network model includes: a heart sound feature extraction module, a convolution block, a maximum pooling layer, a random dropout layer, a channel attention layer, a self-attention layer, an LSTM module, and a first fully connected layer connected in sequence; the heart sound feature extraction module extracts a plurality of heart sound features, and splices and fuses them to form a multidimensional feature matrix as input data of the heart sound classification network model, wherein the heart sound features include Mel-frequency cepstral coefficients, Mel-frequency spectrograms, spectral contrast, and time-frequency features of heart sound signals extracted by short-time Fourier transform.

[0083] Specifically, the collected heart sound data is divided into a training set and a test set. The model is trained using the training set until the deep learning model converges. The trained model is then tested using the test set. The model is tuned based on the test results, and the optimal model is saved.

[0084] Furthermore, the heart sound feature extraction module adopts a variety of audio feature extraction methods, including short-time Fourier transform (STFT), which performs Fourier transform on four types of signals: normal heart sounds, heart murmurs, extra heart sounds, and premature beats in a short time to obtain the distribution information of the signals at different times and frequencies, thereby obtaining the time-frequency characteristics of the four types of signals.

[0085] Obtaining Mel-frequency cepstral coefficients (MFCCs): By performing nonlinear transformation on the spectra of four types of signals (normal heart sounds, heart murmurs, extra heart sounds, and premature beats), the spectral features are converted into cepstral coefficients to capture the spectral envelope characteristics of the signal and reflect the signal's timbre information.

[0086] Obtaining a Mel-scaled Spectrogram: This function simulates the human ear's perception of sounds of different frequencies and converts the spectra of four types of sounds (normal heart sounds, heart murmurs, extra heart sounds, and premature beats) to a Mel-scale frequency scale, highlighting the frequency domain characteristics of the signal.

[0087] Obtain Spectral Contrast: Describes the energy differences between different frequency bands in the spectrum, reflecting the distribution characteristics of the spectrum energy of four categories: normal heart sounds, heart murmurs, extra heart sounds, and premature beats.

[0088] The extracted features are concatenated and fused to form a two-dimensional matrix as the input data of the model.

[0089] Furthermore, a one-dimensional convolutional neural network (Conv1D) was used to process the feature matrix. Conv1D can effectively capture local patterns and dependencies in time series data, preserving more data details. Batch Normalization was then used to improve the training efficiency and performance of deep neural networks.

[0090] Furthermore, the ReLU function with a faster convergence speed is selected as the activation function to enhance the nonlinear expression ability of the data after batch normalization layer processing; the ReLU function can effectively alleviate the gradient disappearance problem and accelerate the training process of the model.

[0091] Furthermore, the maximum pooling layer is used to further filter representative features in the heart sound signal by finding the maximum value within the neighborhood of the pooling filter. The maximum pooling layer can not only reduce the estimated mean shift caused by the signal convolution operation, but also retain more detailed information of the data, thereby improving the robustness of the features.

[0092] Furthermore, a layer of random dropout is used to reduce the structural risk of the model and speed up training. Random dropout forces the model to learn more robust feature representations by randomly discarding some neurons during training, thereby effectively preventing overfitting.

[0093] The above operation is repeated three times to form three convolution blocks, which gradually extract the multi-level features of the heart sound signal.

[0094] Furthermore, the channel attention mechanism is used to further optimize the model's attention to different feature channels. The channel attention mechanism weights each channel of the feature map so that the model can automatically learn the importance of different channel features. Its calculation formula is: , where is the input feature map, for Function that maps the output value to interval, as the weight coefficient of the channel, For the feature map Perform global average pooling operation, the For the feature map Perform a global maximum pooling operation to obtain global information on the channel dimension. It is a multi-layer perceptron that performs nonlinear transformation on the pooled results to learn the dependencies between channels. Finally, the weighted fusion of different channel features is achieved by multiplying the generated weight coefficient with the original feature map.

[0095] Furthermore, a self-attention mechanism is used to capture the dependencies between heart sound signals at different time steps or spatial positions. The self-attention mechanism can dynamically calculate the degree of correlation between each position and all other positions, thus demonstrating powerful performance when processing sequence data. Its calculation formula is: , where 、 and stated are three matrices obtained by linearly transforming the input feature matrix. for The dimensions of the matrix used to The result is scaled to prevent the gradient from disappearing or exploding. The function is used to normalize the calculated attention scores to obtain the attention distribution between each position and other positions; finally, the attention distribution is compared with The matrix is ​​multiplied to obtain the output of the self-attention mechanism, which contains the correlation information between each position in the input sequence.

[0096] Furthermore, two LSTM layers are used to capture the temporal dependencies of heart sound signals. The LSTM layer can effectively model the long-term temporal information in heart sound signals. Each LSTM layer is followed by a layer of random dropout to further reduce the risk of overfitting and enhance the generalization ability of the model.

[0097] Finally, the processed heart sound features are input into the fully connected layer and added The activation function maps the output result to the interval (0,1), indicating the probability distribution of the heart sound channel data belonging to the four categories of normal heart sound, heart murmur, extra heart sound, and premature beat. The calculation formula is: , where The first elements, the is the total number of categories, through the exponential function Map the input value to the positive range to ensure that the output probability value is positive; Sum all the exponential operation results to ensure that the sum of the output probabilities is 1; The output is a probability distribution representing the predicted probability of each category. The model is evaluated and optimized using appropriate evaluation metrics, and the optimal model is saved as the heart sound classification model for normal heart sounds, heart murmurs, extra heart sounds, and premature beats.

[0098] Please refer to Figure 6 In some embodiments, the lung sound classification network model includes: a lung sound feature extraction module, a multi-scale convolution block module, an intra-scale attention module, an inter-scale attention module, a second fully connected layer, an LSTM module, and a third fully connected layer connected in sequence; the lung sound feature extraction module extracts a number of lung sound features and splices them to form a multi-dimensional feature matrix as input data of the lung sound classification network model, wherein the lung sound features include Mel-frequency cepstral coefficients, Mel-frequency spectrograms, zero-crossing rate, and spectral centroid.

[0099] Specifically, the collected lung sound data is divided into a training set and a test set. The model is trained using the training set until the deep learning model converges. The trained model is then tested using the test set. The model is tuned based on the test results, and the optimal model is saved.

[0100] Furthermore, the lung sound feature extraction module extracts multiple features: Mel-frequency cepstral coefficients (MFCCs): used to capture the spectral envelope characteristics of the five types of lung sound signals; Mel-spectrogram: used to extract the frequency domain characteristics of the five types of lung sounds; Zero Crossing Rate: used to reflect the time domain change rate of the five types of lung sound signals, reflecting the fluctuation of the lung sound signals in the time dimension; Spectral Centroid: used to describe the concentrated location of the spectral energy of the five types of lung sounds, helping to understand the distribution trend of lung sound energy in the spectrum. After the features are extracted, a multi-dimensional feature matrix is ​​formed through splicing operations to provide high-dimensional input data for subsequent processing.

[0101] After the multidimensional features are input into the lung sound classification model, the model first extracts features at different scales through a multi-scale convolutional module. Four convolutional blocks of different scales are used, each consisting of five convolutional layers, each of which incorporates a one-dimensional convolutional neural network, batch normalization, and max pooling operations, using the CELU activation function. The convolutional blocks use progressively larger kernel sizes, in the order [10, 8, 6, 4, 2], [20, 16, 12, 8, 4], [30, 24, 18, 12, 6], and [40, 32, 24, 16, 8]. The first convolutional block uses a smaller kernel to capture high-frequency details; the second uses a medium-sized kernel to extract mid-frequency features; the third uses a larger kernel to focus on low-frequency features, extracting long-term trends and global structural information from the lung sound signal; and the fourth uses the largest kernel to capture global features and comprehensively analyze the overall morphology and context of the signal. The features extracted by multi-scale convolution blocks are rich in hierarchy and can provide more comprehensive input for subsequent attention mechanisms and LSTM.

[0102] Next, the multi-scale features are enhanced through the attention module. The attention module first calculates the intra-scale attention of the features of each scale separately, and the calculation formula is , where is the attention mechanism, is the query vector, For the key vector and the is a value vector, for The dimensions of the matrix used to The result is scaled to prevent the gradient from disappearing or exploding. The function is used to normalize the calculated attention scores to obtain the attention weights between each position and other positions. ; Finally, the attention weight and The vectors are multiplied to obtain the output of the attention mechanism, which contains the correlation information between each position in the input sequence.

[0103] Furthermore, the original features are connected through residual and attention-weighted features Fusion , enhance the feature expression ability, and its calculation formula is: , where is the fused feature, is the original feature, is the feature after attention weighting.

[0104] Furthermore, after the intra-scale attention, the model further fuses features of different scales through inter-scale attention; the features fused at scale one are residually connected with the features fused at scale two and the features fused at scale three are residually connected with the features fused at scale four and the features are residually connected and activated, and then respectively input into the attention module, the inter-scale attention scores are calculated and the residual connections are performed to fuse the features, and then the features of all scales are further integrated through the attention mechanism and residual connections to form a unified feature representation.

[0105] Furthermore, the unified feature representation is input into the LSTM layer for sequence modeling; the LSTM layer captures the long-term dependencies in the time series and maps the features to the classification space through the fully connected layer. The Softmax function is used to generate the probability distribution of lung sound channel data belonging to the five categories of normal lung sounds, whistling sounds, moist rales, weakened alveolar breath sounds, and enhanced alveolar breath sounds; the model is evaluated and optimized using appropriate evaluation indicators, and the optimal model is saved as the lung sound classification model for the five categories of normal lung sounds, whistling sounds, moist rales, weakened or disappeared alveolar breath sounds, and enhanced alveolar breath sounds.

[0106] Please refer to Figure 7 In some embodiments, the bowel sound classification network model includes: a bowel sound feature extraction module ResNet50 network module and a multi-head attention mechanism module connected in sequence; the bowel sound feature extraction module extracts several bowel sound features and splices them into a multi-dimensional feature matrix as input data of the lung sound classification network model, wherein the bowel sound features include Mel-frequency cepstral coefficients and Mel-frequency spectrum.

[0107] Specifically, the collected bowel sound data is divided into a training set and a test set. The model is trained using the training set until the deep learning model converges. The trained model is then tested using the test set. The model is tuned based on the test results, and the optimal model is saved.

[0108] Furthermore, the bowel sound feature extraction module extracts multiple features, including Mel-frequency cepstral coefficients (MFCCs), which are used to capture the spectral envelope characteristics of bowel sound signals and extract the frequency domain information of the signal; and Mel-spectrogram, which extracts the time-frequency characteristics of bowel sound signals.

[0109] Furthermore, ResNet50 was selected as the backbone network. ResNet50 is composed of multiple residual blocks stacked together, with a total of 50 layers. The improved ResNet50 network model consists of six parts. The training sample first enters the first part, which includes: a convolution layer with a convolution kernel size of 7×7, a step size of 2, a padding of 3, and out_channels of 64, which is used to preliminarily extract the basic features of the multi-dimensional feature matrix; a batch normalization layer BN (Batch Normalization), which normalizes the output of the convolution layer to speed up the convergence of the model and enhance stability; an activation layer (ReLU activation function), which introduces nonlinear factors to enhance the expressive power of the model; followed by a maximum pooling layer with a convolution kernel size of 3×3, a step size of 2, and a padding of 1, which is used to reduce the dimension of the feature map and reduce the amount of calculation; and finally a multi-head attention mechanism, which highlights important feature information by calculating the correlation weights between different features. The calculation formula of the multi-head attention mechanism is: , where It is a multi-head attention mechanism, 、 and stated are three matrices obtained by linearly transforming the input feature matrix. is the number of heads, indicating the number of times the attention mechanism is executed in parallel. For use The first indivual Attention, the For the general The outputs of the heads are concatenated according to the dimensions. is a learnable weight matrix used to perform a linear transformation on the concatenated result to obtain the final multi-head attention output.

[0110] Next, we proceed to Parts 2 through 5. These four parts primarily consist of residual blocks, which are divided into identity residual blocks (Residual Block I) and convolutional residual blocks (Residual Block II). In each part, the number of convolutional residual blocks is 1, while the number of identity residual blocks varies: 2, 3, 5, and 2, respectively. Both the identity residual block and the convolutional residual block consist of convolutional layers with kernel sizes of 1×1, 3×3, and 1×1, respectively. Each convolutional layer is followed by a batch normalization layer (BN) and a ReLU activation layer (ReLU) to enhance network stability. Part 2 introduces a multi-head attention mechanism after the first residual block to further explore potential relationships between features and enhance feature representation. Part 3 performs feature extraction and transfer according to the conventional residual block structure. Part 4 introduces the multi-head attention mechanism after the residual block structure. Part 5 performs feature extraction and transfer according to the conventional residual block structure. After processing features through a series of residual blocks, a multi-head attention mechanism is used to reweight and filter the features, capturing long-range dependencies between features and enabling the model to more comprehensively understand the input information. Finally, the sixth section comprises an average pooling layer and a fully connected layer. The average pooling layer performs dimensionality reduction on the previously extracted features, converting the feature map into a fixed-length feature vector. The fully connected layer then classifies these feature images and outputs a probability distribution for the bowel sound channel data belonging to one of five categories: normal bowel sounds, active bowel sounds, hyperactive bowel sounds, hyperactive bowel sounds, and absent bowel sounds. The model is evaluated and optimized using appropriate evaluation metrics, and the optimal model is saved as the bowel sound classification model for the five categories: normal bowel sounds, active bowel sounds, hyperactive bowel sounds, hyperactive bowel sounds, and absent bowel sounds.

[0111] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0112] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0113] The above are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's description and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A signal processing method based on multi-physiological sound separation and artificial intelligence, characterized in that the steps include: S1, acquiring multi-channel physiological sound signals of the heart, lungs, and gastrointestinal tract, and performing pre-separation processing on the physiological sound signals to obtain physiological sound components; S2, establishing a BP neural network model, inputting the physiological sound components into the trained BP neural network model to perform component classification and reconstruct the source signal to obtain heart sound signals, lung sound signals, and bowel sound signals; S3, establishing a classification network model, inputting the heart sound signal, the lung sound signal, and the bowel sound signal into the corresponding classification network model respectively for type classification, to obtain the heart sound type, lung sound type, and bowel sound type.

2. The signal processing method based on multi-physiological sound separation and artificial intelligence according to claim 1, characterized in that: The S1 step specifically includes: performing noise reduction on the physiological sound signal; Performing empirical mode decomposition on the denoised physiological sound signal to obtain intrinsic mode function components, and performing Hilbert transform on each of the intrinsic mode function components to calculate the instantaneous frequency; The intrinsic mode function matrix components including the instantaneous frequency are established and analyzed by the independent vector analysis algorithm to obtain the separation component matrix.

3. The signal processing method based on multi-physiological sound separation and artificial intelligence according to claim 2, characterized in that: The S2 step specifically includes: Inputting the independent separated components in the separated component matrix into the trained BP neural network model for component classification, wherein the physiological sound components of the cardiac channel retain the heart sound components, the physiological sound components of the pulmonary channel retain the lung sound components, and the physiological sound components of the gastrointestinal channel retain the bowel sound components; The heart sound component, the lung sound component and the bowel sound component are converted from spectrum signals to time domain signals by inverse short-time Fourier transform to obtain heart sound signals, lung sound signals and bowel sound signals.

4. The signal processing method based on multi-physiological sound separation and artificial intelligence according to claim 1, characterized in that: The training steps of the BP neural network model include: S201, classifying and labeling the physiological sound components, and extracting physiological sound features of the physiological sound components; S202, training the BP neural network model using the physiological sound features and corresponding component labels as training samples; The physiological sound features include: The power ratio of the 10Hz-50Hz frequency band to the 10Hz-2000Hz frequency band, the power ratio of the 50Hz-200Hz frequency band to the 10Hz-2000Hz frequency band, the power ratio of the 200Hz-400Hz frequency band to the 10Hz-2000Hz frequency band, the power ratio of the 400Hz-1000Hz frequency band to the 10Hz-2000Hz frequency band, the power ratio of the 1000Hz-1200Hz frequency band to the 10Hz-2000Hz frequency band, the power ratio of the 1200Hz-2000Hz frequency band to the 10Hz-2000Hz frequency band, the standard deviation of the ventricular contraction cycle interval, and the peak frequency of the power spectral density.

5. The signal processing method based on multiple physiological sound separation and artificial intelligence according to claim 4, characterized in that: The training step of the BP neural network model also includes: S203, assigning the same initial weight to each training sample; S204, training a BP neural network weak classifier based on the weight of the current training sample, adjusting and calculating the weight of the weak classifier to minimize the classification error; S205, increase the weight of the training samples that are misclassified and reduce the weight of the training samples that are correctly classified; S206, determining whether the classification error of the current weak classifier is lower than a preset threshold; if not, returning to step S24 to train the next weak classifier; if so, stopping the training and proceeding to step S207; S207: All weak classifiers and their weights are combined into a strong classifier, and the component classification result is outputted through the strong classifier.

6. A signal processing device based on multi-physiological sound separation and artificial intelligence, characterized in that: include: a pre-separation module, configured to obtain multi-channel physiological sound signals of the heart, lungs, and gastrointestinal tract, and perform pre-separation processing on the physiological sound signals to obtain physiological sound components; a component classification module, configured to establish a BP neural network model, input the physiological sound components into the trained BP neural network model for component classification and reconstruct the source signal to obtain heart sound signals, lung sound signals, and bowel sound signals; The type classification module is used to establish a classification network model, and input the heart sound signal, the lung sound signal and the bowel sound signal into the corresponding classification network model for type classification to obtain the heart sound type, lung sound type and bowel sound type.

7. The signal processing device based on multi-physiological sound separation and artificial intelligence according to claim 6, characterized in that: The type classification module specifically includes: a heart sound classification module, configured to input the heart sound signal into a heart sound classification network model for type classification, and determine the heart sound type using a weighted fusion approach, wherein the heart sound type includes normal heart sound, heart murmur, extra heart sound, and premature beat; a lung sound classification module, configured to input the lung sound signal into a lung sound classification network model for type classification, and determine the lung sound type using a weighted fusion approach, wherein the lung sound types include normal lung sounds, whistles, crackles, weakened alveolar breath sounds, and strengthened alveolar breath sounds; The bowel sound classification module is used to input the bowel sound signal into the bowel sound classification network model for type classification, and determine the bowel sound type by weighted fusion, wherein the bowel sound type includes normal bowel sound, active bowel sound, hyperactive bowel sound, weakened bowel sound, and absent bowel sound.

8. The signal processing device based on multi-physiological sound separation and artificial intelligence according to claim 7, characterized in that: The heart sound classification network model includes: The heart sound feature extraction module, convolution block, maximum pooling layer, random inactivation layer, channel attention layer, self-attention layer, LSTM module, and the first fully connected layer are connected in sequence; The heart sound feature extraction module extracts several heart sound features and splices and fuses them to form a multidimensional feature matrix as input data for the heart sound classification network model. The heart sound features include Mel-frequency cepstral coefficients, Mel-frequency spectrograms, spectral contrast, and time-frequency features of heart sound signals extracted through short-time Fourier transform.

9. The signal processing device based on multi-physiological sound separation and artificial intelligence according to claim 7, characterized in that: The lung sound classification network model includes: The lung sound feature extraction module, multi-scale convolution block module, intra-scale attention module, inter-scale attention module, second fully connected layer, LSTM module, and third fully connected layer are connected in sequence; The lung sound feature extraction module extracts several lung sound features and splices them into a multidimensional feature matrix as input data for the lung sound classification network model. The lung sound features include Mel-frequency cepstral coefficients, Mel-frequency spectrogram, zero-crossing rate, and spectral centroid.

10. The signal processing device based on multi-physiological sound separation and artificial intelligence according to claim 7, characterized in that: The bowel sound classification network model includes: The bowel sound feature extraction module, ResNet50 network module, and multi-head attention mechanism module are connected in sequence; The bowel sound feature extraction module extracts several bowel sound features and splices them into a multidimensional feature matrix as input data for the lung sound classification network model, wherein the bowel sound features include Mel-frequency cepstral coefficients and Mel-frequency spectrograms.

Images